I am currently the co-director of the
Heffner Biomedical Imaging Laboratory, at Columbia University. You can
check our current projects on cardiac US, lung CT and PET images here
Segmentation of ribs on thoracic CT Images
project aims to finely segment rib bones on a large
heterogeneous databases of CT-images from patients and cadavers, toward
the construction of numerical models of the human skeletons for car
manufacture. This project is in collaboration with the LAB
Characterization of pathological process evolution in brain
project focuses on the direct comparison of longitudinal MRI data
sets via non-linear equalization of their histogram with Midway mapping
and statistical analysis of significant differences.
This project is performed in collaboration with E. Mandonnet (Hopital
Lariboisière, France) and O. Salvado (CSIRO, Australia)
Mother Anatomical Modeling from MRI and 3DUS Images
project focuses on generation of accurate and rich anatomical models of
pregnant women from MRI and 3DUS segmentation of fetal structures and
computer graphics tools for the modeling of the maternal envelope.
This project is performed in collaboration with G. Grangé (Hopital Port
Royal, France) and C. Adamsbaum (Hopital Kremlin Bicetre, France).
This project uses the open-source SOFA physics-based modeling
framework. All models a distributed freely to the scientific community.
Variational Segmentation with Multi-phases &
to Thorax PET/CT data
CT data (coronal
PET registered data (coronal view)
- Applications to brain MRI (multi-protocols)
of Internal Gray Matter Nucleii with Fuzzy Spatial
Relations Embedded in a Minimal Surface Framework
Hierarchical Graphs for Anatomical and Funtional Representations of the
Human Anatomy with Pathologies: Application in Oncological neurology.
project is interested in the development of computational tools for
medical practice support. The two approaches focus on information
processing and knowledge representation for medical image
A methodology and software tools are developed with the aim of managing
an electronic patient record after the extraction, integration and
manipulation of patient’s information via segmentation of
anatomical structures on 3D medical images. In this project, the
clinical application deal with the documentation and treatment planning
of brain tumors in neurology, based on magnetic resonance images (MRI)
of the brain for several types of tumors.
List of Papers
Extraction of Cardiac Dynamic Information from Three-Dimensional
Ultrasound with Speckle Tracking
project is interested in the extraction of deformation metrics from
RT3D ultrasound data of the LV. Experiments on disease patients and dog
studies have been performed, showing high sensitivity of the method to
identify abnormal wall motion.
Current developments of
include comparison to Taggued MRI data.
Segmentation of Brain MRIs with Multiphase
Level-Set Deformable Models
project is interested in the application of the multiphase Chan-Vese
segmentation framework to brain MRI to extract white matter, gray
matter and CSF tissues in a fully automated fashion. A study was
publised in "E. Angelini, T. Song, B. Mensh, and A. Laine, Brain MRI
Segmentation with Multiphase Minimal Partitioning: A Comparative Study,
International Journal of Biomedical Imaging, Vol. 2007, Article ID
10526, 2007 (open access).
Current developments for this project
extension of the segmentation frmaework to vectorial data, combining T1
and T2 MRI images, the application to disease cases and the design of
new homogeneity measures, more discreminant.
of coronary arteries on CT images
of Cardiac Function
with Real-Time Three-Dimensional Ultrasound
The long-term goal of this research project is to design and
validate a method of dynamic analysis for recently developed real-time
three dimensional (RT3D) ultrasound technology that would provide
with a new diagnostic instrument to accurately measure cardiac
The proposed development would make available a portable and low cost
which would be easy to administer and less invasive than existing 3D
including high-speed CT, phase contrast and tagged MRI, and equivalent
(and in some cases superior) to the current “gold
standards” for assessing
cardiac function, for example measures of ejection fraction via MUGA.
developed a new method of denoising and contrast enhancement
of cardiac echos tailored to the characteristics of RT3D ultrasound, as
The proposed integration of the time dimension within this process is
innovative in terms of engineering. Existing methods for segmenting 3D
static organs or 2D dynamic echocardiography are based on physical
that are not compatibly with RT3D ultrasound aquisition. Time is
to describe the deformation of the cardiac wall and in the physics of
in the transducer. The proposed research will demonstrate that by
the time dimension in the denoising/enhancement process we can
improve the quality and robustness of clinical measures of importance
access quantitative cardiac function.
Denoising via Thresholding
with Brushlet Expansion
functions were introduced in 1997 by Fran çois
Meyer and Ronald Coifman for the compression of textured images.
functions decompose a signal along oriented textural patterns ("brush
at specific orientation and resolution. They offer an orthogonal
framework and decompose a single frequency in one coefficient. They
provide a complete arbitrary tiling of the Fourier plane. These
make such complex analysis functions superior to Wavelet packets and
functions in terms of analysis compactness and flexibility. We are
in studying the quality of denoising of various "textured" medical
via thresholding of brushlet expansion. Application to multidimensional
textured medical images such as 3D Ultrasound, PET brain scans and MR
for Color and Multi-modality Images
project was part of the Visible
Human Project for the development of the ITK
toolkit sponsored by the National
Library of Medicine.
interested in combining multiscale-relative fuzzy connectedness,
Voronoi diagram classification and three dimensional deformable model
the segmentation of high-resolution multi-channel images. The hybrid
scheme structure combines: (1) Multiscale relative fuzzy connectedness
for extraction of region of interest and texture statistics within the
object (2) Voronoi diagram classification using the statistics
inside the object (3) Three-dimensional deformable model to smooth the
extracted contours and extract 3D object surfaces out of Voronoi
cells. We have tested the hybrid segmentation method on color data from
the Visible human data sets and on other multi modality medical images
such as (T1-PD) registered MRI and cryogenic sections and we have shown
the superiority of the hybrid method over single ones. Segmentation
are more robust and accurate and require limited user intervention with
manual initialization of seed points inside the object to segment.
Segmentation of knee
cartilage using KL classification (1997-1998).
analysis of Electromyography
(EMG) signals for diagnosis of Carpal Tunnel Syndrome (1996).